Exploring the features of OpenCL 2.0

Transcription

1 Exploring the features of OpenCL 2.0 Saoni Mukherjee, Xiang Gong, Leiming Yu, Carter McCardwell, Yash Ukidave, Tuan Dao, Fanny Paravecino, David Kaeli Northeastern University

2 Outline Introduction and evolution of OpenCL OpenCL 2.0- new features Applications used to explore these features Result and analysis

3 OpenCL Programming and runtime framework Executes applications across heterogeneous platforms First version, OpenCL 1.0 was released in 2009 OpenCL 1.0: Basic programming model OpenCL 1.1/1.2: Memory management & fine grain control OpenCL 2.0: Support for emerging hardware capabilities & improved programmability

4 Features Shared Virtual Memory Dynamic Parallelism Generic Address Space Image Support Android Installable Client Driver Extension

5 Features Shared Virtual Memory Dynamic Parallelism Generic Address Space Image Support Android Installable Client Driver Extension

6 Bigger picture Goal: A benchmark and micro benchmark suite with OpenCL 2.0 applications Features that are interesting in HSA and OpenCL 2.0 HPC Mobile/Embedded Big Data 1. Spectral Clustering 2. Connected Component Labeling 3. Graph-based Segmentation 4. Periodic Greens Function 5. Feature Selection and Outlier Detection 6. 2-D Finite Difference Time-Domain 1. N-channel IIR Filtering 2. Multi-channel Noise Filter using FIR Filtering 3. Speech Recognition using Hidden Markov Models 4. AES encryption/ decryption 5. Convolution Neural Network 6. Shallow Water Simulation 7. Color Histogramming 1. Rating System using MapReduce 2. K-means Clustering 3. Page Rank 4. Bayesian Estimation for Adaptive Spam Filtering/Learning 5. Gene Sequencing

7 Bigger picture Goal: A benchmark and micro benchmark suite with OpenCL 2.0 applications Features that are interesting in HSA and OpenCL 2.0 HPC Mobile/Embedded Big Data 1. Spectral Clustering 2. Connected Component Labeling 3. Graph-based Segmentation 4. Periodic Greens Function 5. Feature Selection and Outlier Detection 6. 2-D Finite Difference Time-Domain 1. N-channel IIR Filtering 2. Multi-channel Noise Filter using FIR Filtering 3. Speech Recognition using Hidden Markov Models 4. AES encryption/ decryption 5. Convolution Neural Network 6. Shallow Water Simulation 7. Color Histogramming 1. Rating System using MapReduce 2. K-means Clustering 3. Page Rank 4. Bayesian Estimation for Adaptive Spam Filtering/Learning 5. Gene Sequencing

8 Exploring the benefits of OpenCL 2.0 CyberSecurity: The Advanced Encryption Standard (AES) Adopted by US government for encryption Input as plain text with 256 bit key produces cipher text Blocks running concurrently Our results show that key expansion is faster on CPU than GPU 14 rounds of AES-256 are performed on GPU

9 Exploring the benefits of OpenCL 2.0 Signal Processing: Finite Impulse Response Filtering Impulse Response of finite duration Input: x[1 n] and b[1 N] output: f[x] Number of taps: N = 1024 Synthesized audio stream input Uses weighted reduction - very common parallel operation

10 Exploring the benefits of OpenCL 2.0 Signal Processing: Infinite Impulse Response Filter Less processing power than FIR for same design Decomposed into multiple parallel 2nd-order (real and complex) IIR for performance N 1 number of real poles N 2 number of complex poles Number of channels = 64 FIR coefficient: c 0 = 3.0 Synthesized audio stream input

11 Exploring the benefits of OpenCL 2.0 Statistical Modeling: Hidden Markov Models Probabilistic meaning of hidden states without prior knowledge Targeting isolated word recognition Matrix form used for coalescing and computational efficiency Uses operations including Matrix multiplication Matrix vector Parallel reduction Uses data & thread level parallelism

12 Ongoing OpenCL 2.0 Evaluation Baseline: OpenCL 1.2 GPU model: AMD Radeon R9 290x (reported in paper) Current use: AMD A K Radeon R7, Kaveri APU GPU Architecture: Compute Cores: 12 (4 CPU & 8 GPU) Global Memory: 512 MB Max Clock frequency: 720 MHz GPU Driver: (VM)

13 AES Results Execution time (sec) OpenCL1.2 OpenCL2.0 Optimizations explored: ü SVM X Dynamic Parallelism Unencrypted file size (MB) Input files contain excerpts of a book Input sizes are varied from 1MB to 1,000MB with constant 256 bit key Small benefits from SVM, which grow with input file size Child kernel is memory intensive, inhibiting dynamic parallelism

14 FIR Results Optimizations explored: ü SVM Execution time (sec) OpenCL1.2 OpenCL2.0 Block size FIR is a streaming application with different block sizes Results show that same kernel runs faster in OpenCL 2.0 Consistent benefits from SVM, which grow with input block size

15 IIR Results Optimizations explored: ü SVM X Workgroup function Execution time (msec) OpenCL1.2 OpenCL Block size Interesting feature - parallel reduction Workgroup function is useful for reduction, but did not work well

16 Exploring Workgroup Function further in IIR 0.4 Execution time (sec) HSA 1.0 final + OpenCL1.2 HSA 1.0 final + OpenCL Block size Workgroup function is useful for reduction, but did not work well in OpenCL 2.0 It works better in HSAIL on HSA, but not as good as reduction

17 Hidden Markov Model Results Optimizations explored: ü SVM ü Dynamic Parallelism Execution time (sec) OpenCL1.2 OpenCL2.0 Number of hidden states Updating expected values for each hidden state is an independent operation - perfect for Dynamic Parallelism!

18 K-means Results Data Mining: K-means algorithm Optimizations explored: ü SVM Execution time (sec) Execution time (sec) OpenCL 1.2 OpenCL 2.0 Well known clustering algorithm. K-means with different number of objects, 34 features, 5 clusters Input file contains features and attributes Consistent benefits from SVM Number of objects Number of objects

19 Shallow Water Simulation Results Physics simulation: Shallow Water Engine Optimizations explored: ü SVM Execution time (sec) OpenCL 1.2 OpenCL Number of objects Number of objects Depicts complex behavior of fluids, wave modeling for interactive systems Predicts matters of practical interest, e.g. internal tides in strait of Gibraltar Mathematically and computationally intense, so expensive to do real-time

20 Summary OpenCL 2.0 introduced new features We have explored the benefits of using them with some benchmarks from a variety of domains SVM provides consistent benefits Exploring issues with utilizing the work-group function The benchmark suite will be released Summer 2015

21 Northeastern University Architecture (NUCAR) Group

22 THANK YOU TO OUR SPONSORS